ssg1/coffee-bags-tabular

Coffee Bags — Premium Pricing (Tabular)

34 retail coffee bags described by seven package attributes, with a binary target for whether a bag is premium-priced per ounce. Built for 24-679.

Property	Value
Splits	train (345), validation (5), test (6)
Features	7
Targets	is_premium (binary), price_per_oz (continuous)
Balance (all originals)	17 premium / 17 standard

Purpose

Can package attributes alone predict whether a coffee is expensive per ounce? Every feature is readable off a bag in seconds — no cupping scores or hidden quality metrics.

Exploratory analysis

The median split at $1.08/oz separates the two classes. The groups are contiguous rather than cleanly divided by brand tier — Peet's, Starbucks and Kicking Horse sit just below the boundary, while a few small premium packages sit far above it.

Left: package size pushes unit price down — every bag of 800 g or more falls in the standard class — though plenty of small bags are cheap too, so size alone does not decide the label. Right: how altitude disclosure lines up with the label — see Limitations.

Composition

One row per product SKU.

Column	Type	Role
source_id	string	identifier
parent_id	string	source row; equals source_id for an unaugmented row
brand	string	identifier — not a feature
origin	string	feature — country, or Blend
altitude_masl	int	feature — metres; 0 when unstated
altitude_stated	int	feature — 1 if the package discloses altitude
processing	string	feature — washed / natural / honey / anaerobic / unstated
roast_level	string	feature — light / medium / medium-dark / dark
weight_g	int	feature — net weight in grams
grind	string	feature — whole / ground
price_usd	float	measurement — regular shelf price
price_per_oz	float	target (continuous)
is_premium	int	target (binary)
augmentation	string	provenance — none or the method used
is_augmented	bool	provenance

Collection

Collected September 2026 from bags on hand — current and previously purchased — plus a trip to a nearby grocery store. Where a bag was no longer available, weight and regular shelf price were confirmed against the retailer's current listing.

Prices are regular shelf prices, not sale prices, and retail prices move — this is a September 2026 snapshot. Sampling was spread across roughly half specialty roasters and half supermarket brands, with sizes from 227 g to 2268 g. Both spreads matter: without the price range a median split would only separate expensive from very expensive, and without the size range price_per_oz would be price_usd rescaled. K-cups and pods were excluded.

Preprocessing and labels

price_per_oz = price_usd / (weight_g / 28.349523125), then is_premium = 1 where that exceeds the dataset median of $1.080/oz.

The threshold comes from the data rather than a hand-picked dollar figure, which balances the classes and keeps personal judgement out of the labelling. Across all 34 originals: 17 premium / 17 standard.

Augmentation

14 children per training row (322 total), alternating two methods with seed 24679. Validation and test rows are never used as parents:

1. Additive Gaussian jitter — noise on weight_g, price_usd, and a stated altitude_masl, σ = 5% of each column's standard deviation.
2. Multiplicative scaling — the same columns × a factor drawn uniformly from ±8%.

Label-preserving by construction: categorical columns and altitude_stated are copied exactly, altitude_masl is perturbed only when already non-zero (so "not disclosed" cannot become a fabricated altitude), is_premium is inherited from the parent and never recomputed, and perturbed values are clipped to their declared domains.

price_per_oz is recomputed, so 5 of 322 children (1.6%) land across the median from their inherited label. Every such parent already sat close to the threshold; these are kept as boundary cases.

Splits and intended use

Split	Rows	Contents
train	345	23 training originals + 322 synthetic
validation	5	Held-out originals, unaugmented
test	6	Held-out originals, unaugmented

Originals were split about 70/15/15, stratified on is_premium, before augmentation, so no validation or test bag has a descendant in train. Use the splits as given.

from datasets import load_dataset

ds = load_dataset("ssg1/coffee-bags-tabular")
train, validation, test = ds["train"], ds["validation"], ds["test"]

Intended for coursework and small-scale experiments in tabular classification and regression.

Limitations

• altitude_stated is a strong baseline. Predicting is_premium = altitude_stated scores 88% on the original split. It records whether a roaster chose to print an altitude — a marketing decision that tracks price — not a property of the coffee. Report accuracy against this figure, not against a 50% chance level.
• brand will leak the target. Exclude it from features.
• 34 original samples is small, so validation holds 5 rows and test holds 6. A single misclassification moves test accuracy by roughly 17 points — treat any single score as a rough indication.
• Augmented rows add density, not new information, so training-set estimates are optimistic.
• is_premium is relative — above the median of this sample, not a general definition of premium coffee.
• Blends dominate origin, leaving few rows per single origin country.
• Prices are a September 2026 US snapshot from one regional market.

Ethical notes

The data describe retail products, not people: no personal or sensitive information, no human subjects. Brand names and prices are public retail facts recorded for coursework. Nothing here supports claims about product quality — is_premium is a statement about price per ounce and nothing more.

License

MIT.

AI usage disclosure

• Collection and verification — done by the author; values read from packages and confirmed against retailer listings. No values were generated by an AI model.
• Synthetic rows — deterministic NumPy (seeded Gaussian jitter and uniform scaling), not a generative model. Reproducible from seed 24679.
• Notebook and card — Claude AI helped structure the notebook and draft documentation. Subject, features, target definition, augmentation parameters, and the leakage assessment were decided and verified by the author.